broadcast batching with minimal set of messages needed

3 files changed, 90 insertions, 31 deletions

diff --git a/broadcast/src/batch.rs b/broadcast/src/batch.rs
index d69771d..42cc3b8 100644
--- a/broadcast/src/batch.rs
+++ b/broadcast/src/batch.rs
@@ -50,6 +50,7 @@ impl MessageBatch {
         self.first_added
             .elapsed()
             .saturating_sub(self.max_message_delay)
+            .saturating_sub(Duration::from_millis(10))
     }
 
     pub fn broadcast(
diff --git a/broadcast/src/handler.rs b/broadcast/src/handler.rs
index 2b470fb..e71edef 100644
--- a/broadcast/src/handler.rs
+++ b/broadcast/src/handler.rs
@@ -1,3 +1,4 @@
+use elsa::sync::FrozenMap;
 use smol::{
     future,
     lock::{Mutex, RwLock},
@@ -10,7 +11,11 @@ use std::{
     time::Duration,
 };
 
-use crate::{batch::MessageBatch, msg::*, topology::Topology};
+use crate::{
+    batch::MessageBatch,
+    msg::*,
+    topology::{NodeId, Topology},
+};
 
 use common::{
     msg::{MessageHeader, Output},
@@ -24,8 +29,9 @@ pub struct BroadcastHandler {
     node_id: String,
     seen: RwLock<HashSet<BroadcastTarget>>,
     topology: Topology,
-    batch: Mutex<MessageBatch>,
+    batch: FrozenMap<NodeId, Box<Mutex<MessageBatch>>>,
     pub(crate) output: Output<BroadcastBody>,
+    max_message_delay: Duration,
     attempted_broadcasts: Mutex<HashMap<MessageID, Task<()>>>,
 }
 
@@ -36,12 +42,13 @@ impl Handler for BroadcastHandler {
         let max_message_delay = MAX_STABLE_DELAY_MS / (node_ids.len() as f32).sqrt() as u32;
 
         let this = Arc::new(Self {
+            max_message_delay,
             node_id,
             topology: Topology::dense(node_ids.clone()),
             seen: RwLock::new(HashSet::new()),
             output,
             attempted_broadcasts: Mutex::default(),
-            batch: Mutex::new(MessageBatch::new(max_message_delay)),
+            batch: FrozenMap::new(),
         });
 
         smol::spawn(this.clone().poll_batch()).detach();
@@ -61,7 +68,7 @@ impl Handler for BroadcastHandler {
                     message,
                 } => {
                     future::zip(
-                        self.receive_broadcast(message),
+                        self.receive_broadcast(&header.src, message),
                         self.send_broadcast_ok(&header.src, msg_id),
                     )
                     .await;
@@ -72,7 +79,7 @@ impl Handler for BroadcastHandler {
                     messages,
                 } => {
                     future::zip(
-                        self.receive_broadcast_batch(messages),
+                        self.receive_broadcast_batch(&header.src, messages),
                         self.send_broadcast_ok(&header.src, msg_id),
                     )
                     .await;
@@ -82,14 +89,14 @@ impl Handler for BroadcastHandler {
                     msg_id: None,
                     message,
                 } => {
-                    self.receive_broadcast(message).await;
+                    self.receive_broadcast(&header.src, message).await;
                 }
 
                 BroadcastBody::BroadcastBatch {
                     msg_id: None,
                     messages,
                 } => {
-                    self.receive_broadcast_batch(messages).await;
+                    self.receive_broadcast_batch(&header.src, messages).await;
                 }
 
                 BroadcastBody::Topology { msg_id, topology } => {
@@ -152,49 +159,76 @@ impl BroadcastHandler {
     }
 
     /// Receive a given message, and broadcast it onwards if it is new
-    async fn receive_broadcast(self: &Arc<Self>, message: BroadcastTarget) {
+    async fn receive_broadcast(self: &Arc<Self>, src: &String, message: BroadcastTarget) {
         let new = self.seen.write().await.insert(message);
         if !new {
             return;
         }
 
-        let mut batch = self.batch.lock().await;
-        batch.add(message);
+        for dest in self.topology.targets(&self.node_id, &src).await {
+            let batch_lock = self.get_batch_lock(&dest);
+
+            let mut batch = batch_lock.lock().await;
+            batch.add(message);
+        }
     }
 
-    async fn receive_broadcast_batch(self: &Arc<Self>, message: Vec<BroadcastTarget>) {
-        let mut batch = self.batch.lock().await;
+    async fn receive_broadcast_batch(
+        self: &Arc<Self>,
+        src: &String,
+        message: Vec<BroadcastTarget>,
+    ) {
         let mut seen = self.seen.write().await;
 
         for message in message.into_iter() {
             if seen.insert(message) {
-                batch.add(message);
+                for dest in self.topology.targets(&self.node_id, &src).await {
+                    let batch_lock = self.get_batch_lock(&dest);
+                    let mut batch = batch_lock.lock().await;
+                    batch.add(message);
+                }
             }
         }
     }
 
-    async fn poll_batch(self: Arc<Self>) {
-        loop {
-            let mut batch = self.batch.lock().await;
-            self.do_batch_check(&mut batch).await;
-
-            let time = batch.sleep_time();
-            drop(batch);
+    fn get_batch_lock(&self, key: &String) -> &Mutex<MessageBatch> {
+        match self.batch.get(key) {
+            Some(x) => x,
+            None => {
+                self.batch.insert(
+                    key.clone(),
+                    Box::new(Mutex::new(MessageBatch::new(self.max_message_delay))),
+                );
 
-            Timer::after(time).await;
+                self.batch.get(key).unwrap()
+            }
         }
     }
 
-    async fn do_batch_check(self: &Arc<Self>, batch: &mut MessageBatch) {
-        if batch.should_broadcast() {
+    async fn poll_batch(self: Arc<Self>) {
+        loop {
             let mut tasks = self.attempted_broadcasts.lock().await;
-            for target in self.topology.all_targets(&self.node_id).await {
-                let msg_id = gen_msg_id();
-                let this = self.clone();
-                tasks.insert(msg_id, smol::spawn(batch.broadcast(this, target, msg_id)));
+            let mut min_sleep_time = self.max_message_delay;
+            for key in self.batch.keys_cloned() {
+                let batch_lock = self.batch.get(&key).unwrap();
+                let mut batch = batch_lock.lock().await;
+                if batch.should_broadcast() {
+                    let msg_id = gen_msg_id();
+                    let this = self.clone();
+                    tasks.insert(msg_id, smol::spawn(batch.broadcast(this, key, msg_id)));
+
+                    batch.clear();
+                }
+
+                let batch_sleep_time = batch.sleep_time();
+                if batch_sleep_time < min_sleep_time {
+                    min_sleep_time = batch_sleep_time;
+                }
             }
 
-            batch.clear();
+            drop(tasks);
+
+            Timer::after(min_sleep_time).await;
         }
     }
 }
diff --git a/broadcast/src/topology.rs b/broadcast/src/topology.rs
index 5e16666..bb10466 100644
--- a/broadcast/src/topology.rs
+++ b/broadcast/src/topology.rs
@@ -14,16 +14,40 @@ impl Topology {
         for node_id in node_ids.iter() {
             top.insert(node_id.clone(), node_ids.iter().cloned().collect());
         }
+        Self::filter_desc(&mut top);
 
         Topology(RwLock::new(top))
     }
 
     /// Replace the current topology with a new one.
-    pub async fn replace(&self, new: TopologyDesc) {
+    pub async fn replace(&self, mut new: TopologyDesc) {
+        Self::filter_desc(&mut new);
         *self.0.write().await = new;
     }
 
-    pub async fn all_targets(&self, node_id: &NodeId) -> HashSet<String> {
-        self.0.read().await.get(node_id).unwrap().clone()
+    fn filter_desc(desc: &mut TopologyDesc) {
+        for (node_id, neighbours) in desc.iter_mut() {
+            neighbours.remove(node_id);
+        }
+    }
+
+    /// Get the next targets from the given topology, for a message
+    /// which has travelled across the given path and is now at node_id.
+    pub async fn targets(&self, node_id: &String, last_node_id: &String) -> HashSet<String> {
+        // Ensure we don't keep holding the read lock
+        let topology = self.0.read().await;
+
+        // Get all nodes the last node sent it to
+        let visited = topology.get(last_node_id);
+
+        let neighbours = topology.get(node_id).unwrap();
+        match visited {
+            Some(visited) => neighbours
+                .difference(&visited)
+                .cloned()
+                .filter(|n| n != node_id)
+                .collect(),
+            None => neighbours.clone(),
+        }
     }
 }